modified mewma control scheme for an analytical process …...based on modified mewma control chart...

© 2013. Alpaben K. Patel & Jyoti Divecha. This is a research/review paper, distributed under the terms of the Creative Commons Attribution-Noncommercial 3.0 Unported License http://creativecommons.org/licenses/by-nc/3.0/), permitting all non-commercial use, distribution, and reproduction inany medium, provided the original work is properly cited.

Global Journal of Computer Science and Technology Software & Data Engineering Volume 13 Issue 3 Version 1.0 Year 2013 Type: Double Blind Peer Reviewed International Research Journal Publisher: Global Journals Inc. (USA) Online ISSN: 0975-4172 & Print ISSN: 0975-4350

Modified MEWMA Control Scheme for an Analytical Process Data

By Alpaben K. Patel & Jyoti Divecha Sardar Patel University, Vallabh Vidayanagar

Abstract - This article introduces Multivariate Modified Exponentially Weighted Moving Average (MMOEWMA) control chart, a chart for detecting shifts of all kinds in case of highly first order vector autoregressive VAR (1) process. This chart is based on modified MEWMA control chart statistic which is a correction of MEWMA chart statistic. The performance of MMOEWMA chart is illustrated along with MEWMA chart for a chemical process data. The average run length (ARL) properties of MMOEWMA scheme are derived using Markov Chain approach. Algorithm for the ARL computation and R-program of monitoring MMOEWMA control chart are provided.

Keywords : abrupt change; average run length; MEWMA; multivariate modified EWMA; vector autoregressive.

GJCST-C Classification : E.m


Strictly as per the compliance and regulations of:


Alpaben K. Patel α & Jyoti Divecha σ

Abstract - This article introduces Multivariate Modified Exponentially Weighted Moving Average (MMOEWMA) control chart, a chart for detecting shifts of all kinds in case of highly first order vector autoregressive VAR (1) process . This chart is based on modified MEWMA control chart statistic which is a correction of MEWMA chart statistic. The performance of MMOEWMA chart is illustrated along with MEWMA chart for a chemical process data. The average run length (ARL) properties of MMOEWMA scheme are derived using Markov Chain approach. Algorithm for the ARL computation and R-program of monitoring MMOEWMA control chart are provided. Keywords : abrupt change; average run length; MEWMA; multivariate modified EWMA; vector autoregressive.

I. Introduction

ultivariate statistical process control is often used in chemical and process industries where autocorrelation is most prevalent. Traditional

multivariate statistical process control techniques are based on the assumption that the successive observation vectors are independent. In recent years, due to automation of measurement and data collection systems, a process can be sampled at higher rates, which ultimately leads to autocorrelation. Consequently, when the autocorrelation is present in the data, it can have a serious impact on the performance of classical control charts. This point has been made by numerous authors, including Berthouex, Hunter, and Pallensen (1978), Harris and Ross (1991), Montgomery and Mastrangelo (1991). Runger (1996) has presented a realistic model that generates autocorrelation and cross correlation and provides a useful approach to characterizing process data. The interpretation of these charts: charts based on modeling residuals is not as simple as the authors suggest, and the alternative engineering feedback control methods are often more appropriate with such highly auto correlated data.

This article considers the problem of monitoring the mean vector of a process in which observations can be a highly first order vector autoregressive VAR (1) and propose a control chart called Multivariate Modified EWMA chart. Multivariate Modified EWMA chart as a modification in MEWMA (Lowery et. al, 1992) chart statistic. Multivariate Modified EWMA chart that Author α : Directorate of Economics and Statistics, Sector-18, Gandhinagar-382009. E-mail : [email protected] Author σ : Department of Statistics, Sardar Patel University, Vallabh Vidayanagar-388120.

combines the features of multivariate Shewhart chart (Hotelling, 1947) and MEWMA chart in a simple way and has ability to detect small as well as large shift as soon as possible as required by some industrial processes with high level of first order vector autoregressive data.

MMOEWMA control statistic gives weight to the past observation vectors in slightly different way than MEWMA and each current change is considered with full weight. This corrects MEWMA statistic for suffering from inertia problem. This article discusses the procedures to construct the Multivariate Modified EWMA chart. Simulate the average run length to assess the performance of the chart. The MMOEWMA vector auto correlated control chart is defined in second section and the derivation of upper control limits is kept with Appendix 1. Further, performance of MMOEWMA monitoring scheme is illustrated along with MEWMA scheme for real multivariate chemical process data in third section. ARL properties of MMOEWMA are derived and compared with MEWMA in fourth section. The comparisons reveals that MMOEWMA scheme outperform MEWMA scheme. Computation of ARL values were carried out using Markov chain approach described in Appendix 2.

II. Multivariate Control Charts for Monitoring the Process Mean

Suppose that the p x 1 random vectors Y1, Y2, Y3, … each representing the p quality characteristics to be monitored, are observed over time. These vectors may represent individual observations or sample mean vectors. To study the performance of the various multivariate control charts, it will be assumed that Yn, n = 1, 2, …, are independent multivariate normal random vectors with mean vectors μn, n =1, 2,… For simplicity, it is assumed that each of the random vectors has the known co-variance matrix Σ.

a) The Multivariate Shewhart Control Chart Hotelling’s (1947) introduced a multivariate

control-chart procedure based on his Hotelling’s chi-

square statistics defined as nnn YY 1'2 −Σ=χ . At any nth

stage in process,

nnn YY 1'2 −Σ=χ > h, (1)

signals that a statistically significant shift in the mean has occurred; that is process out-of-control, where h >

M

© 2013 Global Journals Inc. (US)

Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

23

(DDDD DDDD

)Year

013

2C

0 is a specified control limit. Because this procedure is based on only the most recent observation, it is insensitive to small and moderate shifts in the mean vector.

b) The Multivariate EWMA (MEWMA) control chart Lowery et al. (1992) proposed the MEWMA

chart as natural extension of EWMA chart. It is a popular chart used to monitor a process with p quality characteristics for detecting small to moderate shifts. The in-control process mean is assumed without loss of generality to be a vector of zeros, and covariance matrix Σ. The MEWMA control statistic is defined as vectors,

Xn =ΛYn + (I-Λ)Xn-1, n=1,2,…, (2)

where X0 = 0, 1 x p vector and Λ = diag(λ1, λ2,…, λp), 0<λj≤1, j=1,2,…,p.

The MEWMA chart gives an out-of-control signal as soon as

21nT = Xn

' Σxn-1 Xn > h1, (3)

where h1 (>0) is chosen to achieve a specified in control ARL and Σxn is the covariance matrix of Xn given by Σxn = {λ/(2- λ)}Σ, under equality of weights of past observations for all p characteristics; λ1 =λ2=…= λp =λ.

The UCL = 2/11

2/1

)(2

h

− λλ

. If one or more points fall

beyond h1, the process is assumed to be out-of-control. The magnitude of the shift is reflected in the non-centrality parameter μ1

’Σ-1μ1. They conclude that an assignable causes result in a shift in the process mean from μ0 to μ1.

c) MMOEWMA control chart for monitoring the first order vector autoregressive VAR (1) process mean

The MMOEWMA chart as natural extension of Patel and Divecha (2011) proposed Modified EWMA chart. The desirable properties of a multivariate auto

correlated control chart are that it is easy to implement and is effective for detecting shifts of all sizes as per technical specifications. The Multivariate Modified EWMA chart that introduce considers past observations similar to MEWMA scheme and additionally considers past as well as latest change in the process. Let Yn, n = 1,2,…, are sequence of first order auto correlated normal random vectors with mean vector μn, and common covariance matrix Σ. Further it is assumed without loss of generality that the in control process mean vector is μ0 = (0, 0,…, 0)’ = 0.

The MMOEWMA chart statistic is a modification in MEWMA chart statistic. To define MMOEWMA control statistic as vector Xn given by,

Xn= ΛYn + (I- Λ)Xn-1 + (Yn-Yn-1), n≥1, (4)

where X0 is the p-dimensional zero vector and Λ = diag(λ1, λ2,…, λp), 0<λj≤1, j=1,2,…,p. The MMOEWMA chart gives an out-of-control signal as soon as

22nT = Xn

' Σxn-1 Xn > h2, (5)

where h2 (>0) is chosen to achieve a specified in control ARL. If one or more points fall beyond h2, the process is assumed to be out-of-control. Σxn is the covariance

matrix of Xn given by Σxn = Σ

−−

+− λ

λλλ

λ2

)1(22

,

under equality of weights of past observations for all p characteristics; λ1 =λ2=…= λp =λ, and past and current changes. The upper control limit of MMOEWMA chart is,

UCL =2/1

2)1(2

2

−−

+− λ

λλλ

λ 2/12 )(h (discussed in

Appendix 1).

III. Illustration

Table 1 : Temperatures Data

NO Observations Deviated Observations Y1 Y2 Y3 Y1-μ01 Y2-μ02 Y3-μ03

0 92.24 95.56 100.27 0.00 0.00 0.00

1 92.83 95.16 100.77 0.59 -0.40 0.50 … … … … … … … 12 92.84 95.16 100.76 0.60 -0.40 0.49 13 92.84 95.16 100.76 0.60 -0.40 0.49 14 92.84 95.16 100.76 0.60 -0.40 0.49 15 92.84 95.16 100.76 0.60 -0.40 0.49 … … … … … … …

Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

24

(DDDD

)


Year

013

2C


279 92.59 95.10 100.47 0.35 -0.46 0.20

280 92.59 95.10 100.46 0.35 -0.46 0.19 281 92.58 95.10 100.46 0.34 -0.46 0.19

282 92.58 95.10 100.46 0.34 -0.46 0.19

… … … … … … … 588 91.37 95.27 99.73 -0.87 -0.27 -0.54

589 91.37 95.27 99.72 -0.88 -0.29 -0.55

590 91.36 95.27 99.72 -0.88 -0.29 -0.55 591 91.36 95.27 99.72 -0.88 -0.29 -0.55 592 91.35 95.28 99.72 -0.89 -0.28 -0.55

… … … … … … … 998 92.30 96.22 100.16 0.06 0.66 -0.11 999 92.30 96.21 100.16 0.06 0.65 -0.11

… … … … … … … 1200 92.67 95.84 100.72 0.43 0.28 0.45 1201 92.67 95.83 101.95 0.43 0.27 1.68

1202 92.67 95.83 100.73 0.43 0.27 0.46 … … … … … … …

1416 92.55 95.35 100.90 0.31 -0.21 0.63

1417 92.55 95.35 100.90 0.31 -0.21 0.63 … … … … … … …

1434 92.54 95.31 100.87 0.30 -0.26 0.60 1435 92.53 95.30 100.87 0.29 -0.26 0.60

… … … … … … … 1439 92.53 95.29 100.86 0.29 -0.27 0.59

a) MMOEWMA chart for monitoring Multivariate Chemical Process using table 1 and table 2

Table 1 displays the part of measurements on three temperature column taken every minute from a chemical process that is working in control and out of

control situations, abrupt changes and small shifts occur. Here number of variable p=3. Three variables are temperature columns Y1

with mean μ01=92.24, Y2 with

mean μ02=95.56, Y3 with mean μ03= 100.27. To assume

covariance matrix to be

∑

=

)(),(),(),()(),(),(),()(

332313

322212

312111

YVYYCYYCYYCYVYYCYYCYYCYV

=

15.05.0

5.015.0

5.05.01

Upper control limit of MEWMA

UCL = 2/11

2/1

)(2

h

− λλ

=0.882, where λ=0.1,

h1=14.7, and MMOEWMA UCL

=2/1

2)1(2

2

−−

+− λ

λλλ

λ 2/12 )(h =0.839, where

λ=0.1, h2 = 5.417. UCL is used in average run length to

choose appropriate value of decision interval. The MMOEWMA chart gives an out-of-control signal as soon as

22nT = Xn

' Σxn-1 Xn

> h2, h2

= 5.417

and the MEWMA chart gives an out-of-control signal as soon as

21nT = Xn

' Σxn-1 Xn

> h1, h1 = 14.78.

Table 2 shows that, MMOEWMA vector gives best forecasts for the process mean vector; undoubtedly better than the MEWMA prediction barring abrupt change situation (1201st observation). MMOEWMA also detects all the shifts more timely as compared to MEWMA for chemical process data.


Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

25

(DDDD DDDD

)Year

013

2C


Table 2 : Monitoring performance of MEWMA and MMOEWMA for the Chemical Process three variate temperature Data

No. MEWMA vector, λ=0.1, h1 = 14.78

MEWMA statistics

MMOEWMA vector, λ=0.1, h2 = 5.417

MMOEWMA statistics

X1 X2 X3 Tn12 X1 X2 X3 Tn2

2 1 0.06 -0.04 0.05 0.24 0.56 0.46 0.55 2.91 … … … … … … … … … 12 0.43 -0.29 0.36 12.48 0.58 -0.14 0.51 5.19 13 0.44 -0.30 0.37 13.48 0.58 -0.16 0.50 5.47* 14 0.47 -0.31 0.38 14.43 0.58 -0.19 0.50 5.73* 15 0.47 -0.32 0.39 15.30* 0.58 -0.21 0.50 5.98* … … … … … … … … …

279 0.39 -0.46 0.21 15.33* 0.36 -0.48 0.19 5.46* 280 0.38 -0.46 0.21 15.17* 0.36 -0.48 0.19 5.40 281 0.38 -0.46 0.21 15.02* 0.36 -0.48 0.19 5.35 282 0.37 -0.46 0.21 14.87* 0.35 -0.48 0.19 5.30 … … … … … … … … …

588 -0.84 -0.29 -0.52 14.26 -0.86 -0.32 -0.55 5.37 589 -0.84 -0.29 -0.52 14.41 -0.86 -0.32 -0.55 5.42* 590 -0.84 -0.29 -0.53 14.56 -0.87 -0.31 -0.55 5.47* 591 -0.85 -0.29 -0.53 14.72 -0.87 -0.31 -0.55 5.53* 592 -0.85 -0.29 -0.53 14.87* -0.87 -0.31 -0.56 5.59* … … … … … … … … …

998 0.02 0.67 -0.13 14.91* 0.05 0.70 -0.11 5.44* 999 0.03 0.67 -0.13 14.75 0.06 0.70 -0.10 5.38 … … … … … … … … …

1200 0.43 0.29 0.43 4.72 0.46 0.32 0.46 1.91 1201 0.43 0.29 0.55 6.49 1.68 1.54 1.81 29.10* 1202 0.43 0.29 0.54 6.33 0.34 0.19 0.45 1.54

… … … … … … … … … 1416 0.31 -0.18 0.64 14.73 0.31 -0.19 0.63 5.24 1417 0.31 -0.19 0.63 14.79* 0.30 -0.19 0.63 5.25

… … … … … … … … … 1434 0.30 -0.23 0.61 15.35* 0.28 -0.25 0.60 5.41 1435 0.30 -0.23 0.61 15.38* 0.28 -0.25 0.60 5.43*

… … … … … … … … … 1439 0.30 -0.24 0.61 15.48* 0.28 -0.26 0.59 5.46*

In Table 2 observe that, MEWMA control chart

detects shifts on observation 15th to 282nd, 592nd

to 998th, and 1417th

to 1439th. The MMOEWMA control chart detects shifts on observation 13th

to 279th, 589th to 998th,

and 1435th to 1439th. MMOEWMA chart detect abrupt

change at observation 1201st, but MEWMA could not detect it.

Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

26

(DDDD

)


Year

013

2C


Figure 1 and 2 depict MEWMA and MMOEWMA statistics charting.

Figure 1 : MEWWA Control Chart (p=3)

Observe the shoot up bar showing abrupt shift in figure 2 which is completely missing in figure 1.

Figure 2 : MMOEWMA Control Chart (p=3)

b) Properties of MMOEWMA scheme and comparison with MEWMA scheme

All the ARL computations were carried out using Markov chain approach described in Appendix 2. MMOEWMA is the chart for multivariate processes

having autocorrelated observations. However, assuming that MEWMA chart can be applied at least to the residual vectors, to compare the ARL values of MMOEWMA with that of MEWMA having common parameters.


Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

27

(DDDD DDDD

)Year

013

2C


Note that 1201st run has abrupt shift in variable Y3.

Table 3 : Average Run Lengths of MEWMA Charts

ARL Values for MEWMA Charts (p=2, p=3 and p=4) from Lowery et. al. (1992) h1= 10.75 12.34 13.10 14.78 15.16 16.94 Shift P =2 and λ =0.10 p =3 and λ = 0.1 p =4 and λ = 0.1 0.0 501 999 502 1007 497 995 0.5 39.5 51.1 45.6 61.2 52.3 68 1.0 12.1 13.7 13.5 15.3 14.5 16.5 1.5 7.03 7.69 7.66 8.40 8.20 8.99 2.0 4.97 5.39 5.43 5.83 5.79 6.23 2.5 3.90 4.23 4.26 4.54 4.51 4.81 3.0 3.27 3.49 3.52 3.75 3.74 3.97

Table 4 : Average Run Lengths of MMOEWMA Charts

ARL Values for Multivariate Modified EWMA (MMOEWMA) Charts (p=2, p=3 and p=4) h2= 3.93 4.525 4.78 5.417 5.546 6.221 Shift p =2 and λ =0.1 p =3 and λ = 0.1 p =4 and λ = 0.1 0.0 500 1000 500 1000 500 1000 0.5 30.5 40.09 31.52 41.57 32.3 42.77 1.0 10.06 11.63 10.54 12.15 10.89 12.55 1.5 5.69 6.38 6.04 6.73 6.30 7.0 2.0 3.87 4.29 4.16 4.58 4.38 4.80 2.5 2.87 3.17 3.12 3.42 3.31 3.61 3.0 2.02 2.24 2.25 2.46 2.62 2.63

Table 3 to table 6 showed that the control limits for MMOEWMA are quite small than those of MEWMA

as well as 2χ chart for the same in control ARL. For two, three, four variable cases, the smaller out of control ARL imply that MMOEWMA chart performs excellent in

detection of shifts, be it small, moderate or large for every of in control ARLs. MMOEWMA chart with λ1

=λ2=…= λp

=λ

=1 is multivariate Shewhart chart

version for multivariate autocorrelated process.

Table 5 : Average Run Lengths of Multivariate Charts

2χ chart

MEWMA Chart

λ =

0.1

0.2

0.4

0.6

0.8

Shift

h

h1

10.6

8.66

9.65

10.29

10.53

10.58

ARL values for p=2 from Lowery et. al. (1992)

0

200.

200

201

199

200

200

0.5

116.

28.1

35.10

51.9

73.6

95.5

1.0

42.

10.2

10.10

13.2

19.3

28.1

1.5

15.8

6.12

5.50

5.74

7.24

10.3

2.0

6.9

4.41

3.80

3.54

3.86

4.75

2.5

3.5

3.51

2.91

2.55

2.53

2.75

3.0

2.2

2.92

2.42

2.04

1.88

1.91

Table 6 : Average Run Lengths of MMOEWMA Charts

MMOEWMA Chart

λ =

0.1

0.2

0.3

0.4

0.6

0.8

h2=

3.135

3.742

4.223

4.705

5.85

7.561

Shift

ARL values for p=2

0

200

200

200

200

200

200

0.5

21.1

24.76

29.9

29.9

51.6

74.98

1.0

8.12

7.88

8.65

8.65

14.93

23.99

1.5

4.78

4.11

4.06

4.06

5.75

8.92

2.0

3.29

2.64

2.44

2.44

2.92

4.12

2.5

2.45

1.87

1.69

1.69

1.86

2.36

3.0

1.90

1.43

1.32

1.32

1.41

1.63

Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

28

(DDDD

)


Year

013

2C


IV. Conclusion

A simple multivariate control chart for monitoring small as well as large shifts in highly first order vector autoregressive VAR (1) process such as multivariate chemical process is given. It is good method to monitor first order vector autoregressive process in chemical/other industries. Appendix 1 Derivation of the Covariance

Matrix of MMOEWMA Statistic xn

By repeated substitution in equation Xn= ΛYn + (I - Λ)Xn-1 + (Yn-Yn-1), n≥1, it can be shown that

Xn= + 0n Y)(I Λ− +∑

=

Λ−1-n

0j1-j-nj-n

j )Y-(Y)(I (i)

The expectation of Xn gives, E(Xn) = μ , (mean of Yn ).

Lemma 1: If λ1 = λ2 = …= λp = λ, then the expression for Multivariate Modified EWMA statistic

Xn = λ∑−

=

1

0

n

j(1-λ)j Yn-j + (1-λ)n Y0+∑

−

=

1

0

n

j(1-λ)j(Yn-j-Yn-j-1)

Taking expectation on both side,

E(Xn) = λ∑−

=

1

0

n

j(1-λ)j E(Yn-j )+ (1-λ)n E(Y0 )+∑

−

=

1

0

n

j(1-λ)j E(Yn-j-Yn-j-1)

But ])1(1[)]1(1[

])1(1[)1(0

nn

j

nj λ

λλλλλ −−=

−−−−

=−∑=

∴ E (Xn) = 0)1(])1(1[ +−+−− µλµλ nn

= μ

Lemma 2 : If the starting value of process is, X0 = µ0 = Y0 and 0< λ ≤ 1 is a constant. The mean is,

E(Xn) = E((1-λ)Xn-1 + λYn + (Yn - Yn-1)) = µ0.

For p = 2 and n=1, we have,

X1 =

−−

+

−

−+

2021

1011

20

10

2

1

21

11

2

1

1001

00

YYYY

YY

YY

λλ

λλ

=

−+−+−+−+

2021202212

1011101111

)1()1(

YYYYYYYY

λλλλ

=

21

11

XX

say,

So that,

Cov(X1) = ∑X1=

)(

),()(

21

211111

XVXXCovXV

(ii)

Where )( 11XV and )( 21XV are the variance of univariate modified EWMA statistic, and

),( 2111 XXCov = λ1

λ2

σ12

+ (1-

λ1)(1-

λ2) σ12

+

cov( 1011 YY − , 2021 YY − ).

Then as per (ii)

Cov(Xn) = ∑Xn=

)(

),()(

2

211

n

nnn

XVXXCovXV

Now Yn’s are autocorrelated normal with

covariance matrix Σ, so that the (Yn-Yn-1) ‘s (n≥1 ) have covariance matrix 2(I-Rho)Σ with Rho = diag(ρy1, ρy2, …, ρyp). Then, taking λ1 =λ2=…= λp =λ and ρy1, ρy2, …, ρyp →1, as n tends to infinity.

Lemma 3: The variance of univariate Modified EWMA (MOEWMA) control statistic Xn is,

V (Xn)

= V(λ∑−

=

1

0

n

j(1-λ)j

Yn-j

)+ V((1-λ)n

Y0 )+ V(∑−

=

1

0

n

j(1-λ)j(Yn-j-Yn-j-1))


Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

29

(DDDD DDDD

)Year

013

2C


V (Xn) = (1-λ)2n

V(Y0) + ),()1(2)()1( 1

1

0

1

0

12222−−−

−

=

−

=

+−∑ ∑ −+− jnjn

n

j

n

j

jjn

j YYCovYV λλλλ +

)](),[()1(2)()1( 211

1

0

1

0

121

2−−−−−−−

−

=

−

=

+−−− −−−+−−∑ ∑ jnjnjnjn

n

j

n

j

jjnjn

j YYYYCovYYV λλ

( ) ( )∑∑−

=−−−−−

+−

=−−−− −−+−−+

1

011

121

01

2 )(,)1()(,)1(n

jjnjnjn

jn

jjnjnjn

j YYYCovYYYCov λλλλ

Since Yn’s are autocorrelated normal with

variance σ2, the variance of (Yn-Yn-1) (n≥1) is 2222

1 )1(222 σρρσσσ −=−= (small when ρ→1). The weights λ(1-λ)2j

decrease geometrically with the age of sample mean. Suppose Yn’s are correlated to the forward fluctuation (Yn-Yn-1) (n ≥ 1) with common

correlation ρ1 and correlated to the backward fluctuation (Yn+1-Yn), (n≥0) with common correlation ρ2, and forward fluctuation (Yn-Yn-1) are correlated to the backward fluctuation (Yn+1-Yn), (n≥1) with common correlation ρ3, then asymptotic variance for large n,

=)( nXV 22

)2()1(2

)2(ρσ

λλλσ

λλ

−−

+−

+)2(

)1(2 2

λλσρ

−−

+)2(

)1)(1(4 23

λλσλρρ

−−−

+

)2(122)1(

)2()1(22 2

22

1

λσρρλ

λσρρ

−−−

+−− (iii)

In normal autocorrelated process (a) with

3ρ nearly negative half and 1ρ , 2ρ nearly equal and

opposite in sign and being monitored for small shifts, (b) with autocorrelation ρ nearly one (ρ→1) the above

expression (iii), reduces to

=)( nXV 22

)2()1(2

)2(ρσ

λλλσ

λλ

−−

+−

(iv)

Let 2

)2()1(2 ρσ

λλλ

−−

is a small value for high

value of ρ and smallλ , sometimes even negligibly small such that modified EWMA limits equal EWMA limits.

Therefore, V(Xn) =

)2( λ

λ−

+)2()1(2

λλλ

−−

σ2. (v)

Therefore, Multivariate MOEWMA covariance from equation (iv, v) becomes

)( 1nXV =

)2( λ

λ−

+)2()1(2

λλλ

−−

Σ = )( 2nXV = ),( 21 nn XXCov .

In general the best approximation of covariance matrix of the MMOEWMA p-variable vectors is given by,

∑ =Xn

Σ−−

+Σ−

ρλλλ

λλ

)2()1(2

)2( (vi)

The UCL of MMOEWMA control chart is

UCL = ( ) 2/12

2/1

)2()1(2

)2(h

−−

+− λ

λλλ

λ (vii)

The MMOEWMA chart gives an out-of-control signal as soon as

22nT = Xn

'

Σxn-1

Xn

> h2,

(viii)

Where h2

(>0) is chosen to achieve a specified

in control ARL. If one or more points fall beyond h2, the

process is assumed to be out-of-control. The magnitude of the shift is reflected in the non-centrality parameter μ1

’Σ-1μ1.

We conclude that an assignable causes result in

a shift in the process mean from μ0

to μ1.

Appendix 2 RL Computation for

MMOEWMA Scheme using Markov Chain

Approach

Following Runger and Prabhu (1996) and Molnau et al. (2001) the Markov chain approach of ARL for MMOEWMA has been derived. Different choices of λ

(weighting factor), h2

(decision value), and p (number of

variable) are considered.

In and out of-Control Case

For the in or out control case, the ARL analysis can be simplified as a one dimensional Markov chain.

Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

30

(DDDD

)


Year

013

2C


A

To approximate nX , we partition the control region

into m+1 transient states, each of width g = )12(

2+m

UCL.

In this case the two dimensional range of Xn is represented by the X1 and X2 axes, and the states used

for the Markov chain are assumed as circular rings. Because Yn has a spherical distribution, the probability of transitioning from state i to state j, denoted as p(i, j), depends only on the radii of states i and j. For i = 0,1,2,…,m and j not equal to zero.

p(i,j) = P (dn in state j | dn-1 in state i)

= P [ (j-0.5) g < )()1( 11 −− −+−+ nnnn YYXY λλ < (j+0.5) g | dn-1 = ig].

Given that dn-1 = ig, Xn-1is distributed as igU, and j =0,1,2,…,m

p(i,j) = P [ (j-0.5) g < )()1( 1−−+−+ nnn YYigUY λλ < (j+0.5) g | dn-1 = ig].

Let e denote the p component unit vector e’ =(1,0,0,…,0). According to Runger and Prabhu (1996) Yn and U are independent spherical random variables,

without loss of generality it can assume that U is identity equal to e to obtain

p(i,j) = P [ {(j-0.5) g }/ λ< λλ /)]()1[( 1−−+−+ nnn YYigeY < {(j+0.5) g} / λ].

Let ),(2 cpχ denote a non central chi square random variable with p degrees of freedom and non centrality parameter c. Then we have For j not equal to zero (j ≠ 0),

p(i,j) = P

+−2

222

2

22 )5.0(),()5.0(λ

χλ

gjcpgj ,

Where c = [{(1-λ) i g / λ}+d]2, degree of freedom is p , d is the shift in mean vector.

For the case where j =0, we have

p(i,0) = P [ χ2

(p,c) < {(0.5)2

g2

/ λ2

}].

For any control chart that is approximated by a

Markov chain, the run length performance can be determined from the transition probability matrix. Assume that a Markov chain has s states (see Brook and Evans (1972)). The transition probability matrix contains the transition probabilities for moving from state to state. Let this s x s matrix of transition probabilities be presented as P, where the process mean vector is such that the non centrality parameter is δ. Let the s x 1 vector q

designate the starting state of

the Markov chain. The vector q will have a one in the

component corresponding to the starting state and zeros in all of the other components. The zero state ARL of a scheme modeled as a Markov chain represented by

ARL=q'(I-P)-1

1. (ix)

Steps of ARL Computation for MMOEWMA

Step-1 Choose the parameter λ (Weighting factor), h2 (decision value), p (number of variable), and shift in mean vector d. Step-2 The upper control limit of MMOEWMA chart is,

UCL =2/1

2)1(2

2

−−

+− λ

λλλ

λ 2/12 )(h .

Step-3 Choose the number of states m.

Step-4 Compute width g =)12(

2+m

UCL.

Step-5 ),(2 cpχ denotes a non central chi square random variable with p degrees of freedom and non centrality parameter c.

Step-6 Non centrality parameter, c = [ {(1-λ) i g / λ}+d]2, degree of freedom is p , d is the shift in mean vector.

Step-7 For j not equal to zero (j ≠ 0),

p(i,j) = P

+−2

222

2

22 )5.0(),()5.0(λ

χλ

gjcpgj .

Step-8 For the case where j =0, we have

p(i,0) = P [ χ2 (p,c) < {(0.5)2 g2 / λ2 }].

Step-9 Adjust the t. p.m.(Ra) such that row sums are unity.

Step-10 Compute u = [I -

R]-1

1

Step-11 Compute q = Ra’* I


Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

31

(DDDD DDDD

)Year

013

2C


Step-12 ARL = q’ * u, OR ARL = q'[I - R]-1 1

References références referencia

1.

Berthouex, P.M., W.G. Hunter,

and L. Pallesen (1978) Monitoring Sewage Treatment Plants: Some quality control aspects. Journal of Quality Technology, Vol.

10.

2.

Brook, D., and Evans, D.A.

(1972) An Approach to the Probability Distribution

of CUSUM Run Lengths. Biometrika, Vol.59, pp.

539-549.

3.

Harris, T.J., and W.H.

Ross (1991) Statistical Process Control Procedures for Correlated Observations. Canadian Journal of Chemical Engineering, Vol.

69.

4.

Hotelling, H. (1947) Multivariate Quality Control, Techniques of Statistical Analysis, edited by Eisenhart, Hastay, and Wallis, McGraw-Hill, New York.

5.

Lowry,

C.A., Woodall,

W.H., Champ, C.W. and Rigdon,

S.E. (1992) A multivariate exponentially weighted moving average control chart. Technoetrics, Vol. 34(1), pp.

46-53.

6.

Molnau, W.E., Runger, G.C., Montgomery, D.C., Skinner, K.R. and Loendo, E.N. (2001) A Program

for ARL Calculation for Multivariate EWMA Charts. Journal of Quality Technology, Vol. 33(4).

7. Montgomery, D.C., and C.M. Mastrangelo (1991) Some Statistical Process Control methods for Autocorrelated data (with discussion). Journal of Quality Technology. Vol.26.

8. Patel, A.K. and Divecha, J. (2011) Modified exponentially weighted moving average (EWMA) control chart for an analytical process data, Journal of Chemical Engineering and Material Science, Vol. 2(1), 12-20.

9. Runger, G.C. and Prabhu, S.S. (1996) A Markov Chain Model for the Multivariate Exponentially Weighted Moving Averages Control Chart. Journal of American Statistical Association, Vol. 91(436), 1701-1706, Theory and Methods.

10. Runger, G.C. (1996) Multivariate statistical process control for autocorrelated process. International Journal of production research, Vol. 34, pp. 1715-1724, 1996.

11. R Language: Copyright 2004, The R Foundation for Statistical Computing Version 2.0.1 (2004-11-15), ISBN 3-900051-07-0.

R-Program for monitoring Modified Multivariate Exponentially Weighted Moving Average Control Chart

## Multivariate MOEWMA (MMOEWMA) ## Three Temperature Data p<-3 X<-read.table("Rprogram/Temp3.txt",header=TRUE) X ##Temperature T3=X1,T11=X2,T21=X3 X1<-as.matrix(X[1:1439,1]) X1 x1<-ts(X1) ar1<-arima(x1,order=c(1,0,1)) a1<-mean(X1) a1<-92.24 ##a1=92.24 X2<-as.matrix(X[1:1439,2]) X2 x2<-ts(X2) ar2<-arima(x2,order=c(1,0,1)) a2<-mean(X2) a2<-95.56 ## a2= 95.56 X3<-as.matrix(X[1:1439,3]) X3 x3<-ts(X3) ar3<-arima(x3,order=c(1,0,1)) a3<-mean(X3) a3<-100.27 ## a3=100.27 ## if we take unit variances and Correlation=0.5 cmat <- matrix(c(1,0.5,0.5,0.5,1,0.5,0.5,0.5,1), nrow = 3, ncol=3, byrow=TRUE) cmat cmat1<-solve(cmat) cmat1 m<-1439 ## Exponential Weight r, 0<r<=1 r<-0.10

Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

32

(DDDD

)


Year

013

2C


## Difference Variance ##k<-0.095 k<-(2*r*(1-r))/(2-r) Xn<-matrix(0,m,3) for(i in 1:m) { for(j in 1:3) { Xn[i,1]<-X[i,1]-a1 Xn[i,2]<-X[i,2]-a2 Xn[i,3]<-X[i,3]-a3 }} round(Xn,2) ## MMOEWMA Vector Zi= rXi+(1-r)Zi-1+(Xi-Xi-1), Z0=M0=X0=0 ##Asymptotic Sz={(r/(2-r))+k}s Sz<-{(r/(2-r))+k}*cmat Si<-solve(Sz) Si Zi<-matrix(0,m,3) Z1<-0 Z2<-0 Z3<-0 X0<-0 for(i in 1:m) { for(j in 1:3) { Zi[i,1]<-r%*%Xn[i,1]+(1-r)*Z1+(Xn[i,j]-X0) Zi[i,2]<-r%*%Xn[i,2]+(1-r)*Z2+(Xn[i,j]-X0) Zi[i,3]<-r%*%Xn[i,3]+(1-r)*Z3+(Xn[i,j]-X0) if(i>1) { Zi[i,1]<-r%*%Xn[i,1]+(1-r)%*%Zi[i-1,1]+(Xn[i,j]-Xn[i-1,j]) Zi[i,2]<-r%*%Xn[i,2]+(1-r)%*%Zi[i-1,2]+(Xn[i,j]-Xn[i-1,j]) Zi[i,3]<-r%*%Xn[i,3]+(1-r)%*%Zi[i-1,3]+(Xn[i,j]-Xn[i-1,j]) } } } round(Zi,2) Tn<-matrix(0,m,1) t1<-matrix(0,1,3) s1<-matrix(0,3,3) for(i in 1:m) { Tn[i,]<-t(Zi[i,])%*%Si[,]%*%Zi[i,] } round(Tn,2) h4<-matrix(5.417,m,1) shift<-matrix(0,m,1) for(i in 1:m) { if(Tn[i]>h4[i]) shift[i]<-1 else shift[i]<-0 } shift n1<-matrix(0,m,1) for(i in 1:m) { n1[i]<-i} n1 plot(n1,Tn,type="l",xlab="Observations",ylab="Tn^2",main="MMOEWMA,p=3") lines(h4) ##End of program


Globa

l Jo

urna

l of C

ompu

ter Sc

ienc

e an

d Te

chno

logy

V

olum

e XIII

Issu

e III

Versio

n I

33

(DDDD DDDD

)Year

013

2C


Global Journals Inc. (US)

Guidelines Handbook

www.GlobalJournals.org

modified mewma control scheme for an analytical process …...based on modified mewma control chart...

Documents